A method, system, terminal, and storage medium for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning.

Abstract

This invention relates to the field of spatiotemporal data processing technology, and discloses a method, system, terminal, and storage medium for predicting multi-dimensional urban indicators based on semi-supervised learning with small samples. The method includes: extracting visual, textual, and structural features of urban space using a three-branch encoder, and generating a unified urban region representation vector using a multimodal fusion mechanism; subsequently, constructing a semi-supervised collaborative computing framework that integrates a heterogeneous neural network model and an ensemble tree model as a regressor; and, under conditions of a small number of labeled samples, using an alternating pseudo-label generation and four-fold pseudo-label selection mechanism, performing semi-supervised augmentation training on the model with massive amounts of unlabeled data to improve the accuracy of spatiotemporal data labeling. This invention significantly reduces the dependence on labeled data and achieves high-precision labeling prediction across multiple dimensions of urban indicator calculation.

Description

Technical Field

[0001] This invention relates to the field of spatiotemporal data processing technology, and in particular to a method, system, terminal, and computer-readable storage medium for predicting small-sample multi-dimensional urban indicators based on semi-supervised learning. Background Technology

[0002] With the acceleration of global urbanization, spatiotemporal big data such as satellite remote sensing, street view data, points of interest, and movement trajectories are experiencing explosive growth, providing a rich data foundation for multidimensional urban indicator prediction and calculation. However, in actual urban indicator calculation tasks, obtaining high-quality target indicator labels is extremely costly, usually relying on large-scale population censuses or expensive sensor deployments. This situation of "massive amounts of unlabeled data and extremely limited labeled samples" poses a serious challenge of sample scarcity for geospatial artificial intelligence models in the field of urban indicator calculation, hindering seamless urban health checks and monitoring.

[0003] Semi-supervised learning provides an effective technical approach to alleviate sample scarcity and reduce labeling costs by uncovering the underlying distribution patterns in massive amounts of unlabeled data. However, in the field of urban computing, existing methods have not fully explored the potential of semi-supervised learning: most mainstream models are based on supervised learning frameworks, and their performance is highly dependent on large-scale labeled samples, making them prone to overfitting in small-sample scenarios; existing methods often lack a systematic and deep integration of semi-supervised learning mechanisms with the characteristics of urban spatial data, resulting in the accuracy and stability of urban indicator calculations under conditions of scarce labeled samples still failing to meet the needs of practical applications.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for predicting urban indicators using small samples based on semi-supervised learning. This aims to solve the problem that the prediction of urban indicators in the prior art relies heavily on large-scale labeled samples, resulting in low accuracy and stability of urban indicator calculation under conditions of scarce labeled samples.

[0006] To achieve the above objectives, this invention provides a method for predicting small-sample multi-dimensional urban indicators based on semi-supervised learning. The method includes the following steps: Multiple modal data of the target area and multiple target city indicator data are obtained, and all the modal data are mapped to spatial units to obtain a modal dataset, wherein the target city indicator data is small sample data and is used as labels for training; The modal dataset is input into a deep feature encoder for feature extraction, outputting physical features, functional features, and structural features. The physical features, functional features, and structural features are mapped to a unified feature space to calculate the correlation weights of the physical features, functional features, and structural features. The physical features, functional features, and structural features are then fused using all the correlation weights to obtain a city spatial representation vector. The city spatial representation vector is used to construct labeled and unlabeled datasets. The constructed deep neural network model and deep ensemble tree model are trained iteratively using the labeled dataset and the unlabeled dataset. The unlabeled dataset is input into the current deep neural network model and the deep ensemble tree model, and the first pseudo-label set and the second pseudo-label set are output. Construct multiple reset confidence verification constraints, use all the confidence verification constraints to filter the target pseudo-label sample set from the first pseudo-label set and the second pseudo-label set, and add multiple pseudo-labels from the target pseudo-label sample set to the labeled dataset to obtain the target labeled dataset; The current deep ensemble tree model is trained using the target labeled dataset, and the deep neural network model is trained in the next round of training until a target prediction model is obtained. The target prediction model is then used to predict the target region to be tested, and multiple indicator prediction results are output.

[0007] Optionally, the method for predicting small-sample multi-dimensional urban indicators based on semi-supervised learning, wherein acquiring multiple modal data of the target area and multiple target urban indicator data, and mapping all the modal data to spatial units to obtain a modal dataset, specifically includes: Acquire multiple modal data of the target region, and project each modal data into the constructed target specification network; The target specification network is queried by a spatial query algorithm, and all modal data in each unit are defined as a modal dataset.

[0008] Optionally, in the semi-supervised learning-based small-sample multi-dimensional urban indicator prediction method, the deep feature encoder includes: a visual environment encoding branch, a text semantic encoding branch, and a spatial structure encoding branch. The modal dataset includes: image dataset, text dataset, and structure dataset; The process of inputting the modality dataset into a deep feature encoder for feature extraction, and outputting physical features, functional features, and structural features, specifically includes: The image dataset is input into the visual environment coding branch, various image data are filtered out, all image data are divided into multiple image blocks, all image blocks are mapped into high-dimensional embedding vectors, the position encoding of each cell is introduced into the high-dimensional embedding vectors, and the global features of the high-dimensional embedding vectors are extracted using a multi-layer attention mechanism to output physical features. The text dataset is input into the text semantic encoding branch, various types of text data are filtered out, all word units in the text data are mapped into comprehensive word vectors, and the comprehensive word vectors are analyzed using a bidirectional encoding structure to output functional features; The structure dataset is input into the spatial structure encoding branch, various spatial data are filtered out, a graph structure is constructed based on all the spatial data, and the graph structure is analyzed using graph convolutional network and graph attention network respectively, and structural features are output.

[0009] Optionally, the method for predicting small-sample multi-dimensional urban indicators based on semi-supervised learning, wherein mapping the physical features, functional features, and structural features to a unified feature space to calculate the correlation weights of the physical features, functional features, and structural features, and fusing the physical features, functional features, and structural features using all the correlation weights to obtain an urban spatial representation vector, and constructing labeled and unlabeled datasets using the urban spatial representation vector, specifically includes: The physical features, functional features, and structural features are mapped to a unified feature space to obtain the corresponding transformation features; A certain transformation feature is randomly defined as the query vector, and the other two transformation features are defined as the key vector and value vector, respectively, to calculate the relevance weights between different transformation features: ; in, This indicates that functional characteristics are modeled using physical features. Indicates physical characteristics, The function represents the feature, and softmax represents the probability transformation function. This represents a query matrix transformed by physical and functional characteristics. The bond matrix represents the transformation between physical and functional characteristics. Indicates transpose. This represents a value matrix transformed by physical and functional characteristics. Indicates the feature dimension; In the adaptive gated fusion layer, all modeled feature representations are concatenated along the channel dimension, and each feature representation is fused according to all the aforementioned relevance weights to output a city spatial representation vector: ; in, Represents the urban spatial representation vector. This indicates an adaptive gating fusion layer. This represents the adaptive gated fusion function. This represents the physical feature representation after modeling. This represents the functional features after modeling. This represents the structural features after modeling. Indicates to , and Perform fusion processing; An initial labeled dataset is constructed using the urban spatial representation vector and the label, and the unlabeled data is constructed into an unlabeled dataset.

[0010] Optionally, the method for predicting small-sample multi-dimensional urban indicators based on semi-supervised learning, wherein the step of iteratively training the constructed deep neural network model and deep ensemble tree model using the labeled dataset and the unlabeled dataset, inputting the unlabeled dataset into the current deep neural network model and the deep ensemble tree model, and outputting a first pseudo-label set and a second pseudo-label set, specifically includes: Constructing the initial deep neural network model and the initial deep ensemble tree model: ; ; in, This represents the initial deep neural network model. This represents the initial deep ensemble tree model. This represents the prediction error loss function. Represents a deep neural network model. Represents a deep ensemble tree model. Represents the first in the labeled dataset Urban spatial representation vector of each sample Indicates the first The label value of the target city indicator for each sample. This represents a labeled dataset. This indicates determining the optimal parameter settings for the deep neural network model. This indicates the optimal parameter settings for the deep ensemble tree model. In the current training epoch, the unlabeled dataset is input into the initial deep neural network model and the initial deep ensemble tree model for prediction, respectively, and outputs a first pseudo-label set and a second pseudo-label set: ; in, and These represent the first set of pseudo-tags and the second set of pseudo-tags, respectively. and They represent the first Deep neural network models and deep ensemble tree models trained in multiple iterations. Indicates the first in the unlabeled dataset One sample.

[0011] Optionally, in the semi-supervised learning-based small-sample multi-dimensional urban indicator prediction method, the confidence verification constraints include: stability constraints, model consistency constraints, distribution rationality constraints, and neighborhood consistency constraints. The step of constructing multiple confidence verification constraints, using all the confidence verification constraints to filter the target pseudo-label sample set from the first pseudo-label set and the second pseudo-label set, and adding multiple pseudo-labels from the target pseudo-label sample set to the labeled dataset to obtain the target labeled dataset, specifically includes: For the current training epoch, calculate the stability constraints of the deep neural network model under multiple random perturbations: ; in, Indicates the first pseudo-label in the set. Stability constraints on the prediction variance of a sample under multiple random perturbations. Indicates the first The first set of pseudo-labels under random perturbation Indicates random perturbation; The model consistency constraints between the current deep neural network model and the current deep ensemble tree model are calculated using the first pseudo-label set and the second pseudo-label set: ; in, This represents the absolute value of the difference between the pseudo-labels output by the current deep neural network model and the current deep ensemble tree model. Indicates the consistency threshold; Obtain the statistical distribution characteristics of the target metrics in the labeled dataset, and use these characteristics to constrain the value range of the pseudo-labels: ; in, Indicates the first Pseudo-labels for each sample This represents the lower quantile of the labeled sample distribution. Indicates the high quantile of the labeled sample distribution; All samples are identified in the dual neighborhoods of the target unit's geographic space and feature space, resulting in a dual neighborhood set. The pseudo-labels of this dual neighborhood set maintain consistency with the mean of the known labels within the neighborhood. ; in, This represents a neighborhood consistency constraint. Indicates the first The neighborhood set of each sample Represents the first in the neighborhood set One sample, Represents the first in the neighborhood set The label corresponding to each sample Indicates the neighborhood consistency threshold; Based on the stability constraint, the model consistency constraint, the distribution rationality constraint, and the neighborhood consistency constraint, multiple pseudo-label samples are selected from the pseudo-label set to obtain the target pseudo-label sample set; The target pseudo-label sample set, the first pseudo-label set, and the pseudo-labels of the second pseudo-label are added to the annotation dataset to obtain the target annotation dataset: ; in, Indicates the first The labeled dataset for each training round. Indicates the first The labeled dataset for each training round. This represents the target pseudo-label sample set.

[0012] Optionally, the method for predicting small-sample multi-dimensional urban indicators based on semi-supervised learning, wherein the current deep ensemble tree model is trained using the target-labeled dataset, and the next deep neural network model is trained in the next training round until a target prediction model is obtained, and the target prediction model is used to predict the target area to be tested, outputting multiple indicator prediction results, specifically including: The current deep ensemble tree model is trained using the target-labeled dataset to update its parameters, resulting in the deep ensemble tree model for the next training epoch. ; in, Indicates the first A deep ensemble tree model with training epochs, Represents a deep ensemble tree model. Indicates the first The training dataset for each training round. This represents the urban spatial representation vector of a sample in the target labeled dataset. This represents the label values ​​of samples in the target labeled dataset. This represents the prediction error loss function; The deep ensemble tree model from the next training round is used to assist the deep neural network model in the current round in iterative training until the target prediction model is obtained. The target prediction model is used to predict the unlabeled dataset and output the prediction results of multiple indicators for the target region.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a small-sample multi-dimensional urban indicator prediction system based on semi-supervised learning, wherein the small-sample multi-dimensional urban indicator prediction system based on semi-supervised learning includes: The data acquisition module is used to acquire multiple modal data of the target area and multiple target city indicator data, and map all the modal data to spatial units to obtain a modal dataset, wherein the target city indicator data is small sample data and is used as labels to participate in training; The multimodal fusion module is used to input the modal dataset into a deep feature encoder for feature extraction, output physical features, functional features, and structural features, and map the physical features, functional features, and structural features to a unified feature space to calculate the correlation weights of the physical features, functional features, and structural features. The module then uses all the correlation weights to fuse the physical features, functional features, and structural features to obtain a city spatial representation vector. The module uses the city spatial representation vector to construct labeled and unlabeled datasets. The pseudo-label generation module is used to perform alternating iterative training on the constructed deep neural network model and deep ensemble tree model using the labeled dataset and the unlabeled dataset, inputting the unlabeled dataset into the current deep neural network model and the deep ensemble tree model, and outputting a first pseudo-label set and a second pseudo-label set; The confidence verification module is used to construct multiple reset confidence verification constraints, use all the confidence verification constraints to filter the target pseudo-label sample set from the first pseudo-label set and the second pseudo-label set, and add multiple pseudo-labels from the target pseudo-label sample set to the labeled dataset to obtain the target labeled dataset; The task output module is used to train the current deep ensemble tree model using the target annotation dataset, train the deep neural network model in the next round of training, until a target prediction model is obtained, and use the target prediction model to predict the target region to be tested, and output prediction results of multiple indicators.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a small sample multi-dimensional urban indicator prediction program based on semi-supervised learning stored in the memory and executable on the processor. When the small sample multi-dimensional urban indicator prediction program based on semi-supervised learning is executed by the processor, it implements the steps of the small sample multi-dimensional urban indicator prediction method based on semi-supervised learning as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a small sample multi-dimensional urban indicator prediction program based on semi-supervised learning, and when the small sample multi-dimensional urban indicator prediction program based on semi-supervised learning is executed by a processor, it implements the steps of the small sample multi-dimensional urban indicator prediction method based on semi-supervised learning as described above.

[0016] In this invention, multiple modal data of a target region and multiple target city indicator data are acquired. All modal data are mapped to spatial units to obtain a modal dataset. The target city indicator data are small sample data used as labels for training. The modal dataset is input into a deep feature encoder for feature extraction, outputting physical features, functional features, and structural features. The physical features, functional features, and structural features are then fused to obtain an unlabeled dataset. The labeled dataset and the unlabeled dataset are used to iteratively train a constructed deep neural network and a deep ensemble tree model. The dataset is input to the current deep neural network and the deep ensemble tree model, and outputs a corresponding set of pseudo-labels. Multiple confidence verification constraints are constructed, and multiple pseudo-label samples are selected from the pseudo-label set using all the confidence verification constraints. The labeled dataset is updated using all the pseudo-label samples to obtain the target labeled dataset. The current deep ensemble tree model is trained using the target labeled dataset, and the deep neural network is trained in the next training round until a target prediction model is obtained. The target prediction model is then used to predict the target area to be tested, outputting prediction results for multiple indicators. This invention significantly reduces the dependence on labeled data and achieves high-precision labeled prediction across multiple city indicator calculation dimensions. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the small-sample multi-dimensional urban indicator prediction method based on semi-supervised learning of the present invention; Figure 2 This is a flowchart of a preferred embodiment of the present invention's method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning; Figure 3 This is a structural diagram of a preferred embodiment of the small-sample multi-dimensional urban indicator prediction system based on semi-supervised learning of the present invention; Figure 4 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] The preferred embodiment of the present invention describes a small-sample, multi-dimensional urban indicator prediction method based on semi-supervised learning, such as... Figure 1 As shown, the method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning includes the following steps: Step S10: Obtain multiple modal data of the target area and multiple target city indicator data, and map all the modal data to spatial units to obtain a modal dataset, wherein the target city indicator data is small sample data and is used as labels for training.

[0020] The target city indicator data is only available in some spatial units and is not fully covered (e.g., air pollution indicators are only available in units with monitoring stations). What this invention aims to achieve is to predict city indicators that are fully covered across the entire region using a small amount of labeled data (i.e., small sample) and publicly available, low-cost multimodal data.

[0021] Specifically, multiple modal data of the target region are acquired, and each modal data is projected into the constructed target specification network. The target specification network is queried by a spatial query algorithm, and all modal data in each unit are defined as a modal dataset.

[0022] In the embodiments disclosed in this invention, spatial unit subdivision and multimodal data alignment are performed first, thus establishing a unified urban spatial research benchmark.

[0023] Based on the scale of the target area (or target city), target network units (500 meters or 1000 meters) are selected as discrete research units, and a globally unique identifier is defined for each target network unit. On this basis, all modal data (such as street view images, remote sensing images, regions of interest, social media, road networks and trajectory data) within the target area can be mapped onto the target specification network constructed from all target grid units.

[0024] Specifically, a coordinate transformation algorithm is used to project the geographic coordinates of these modal data into the same geographic coordinate system. Then, based on the geographic boundaries of each target network unit, spatial query operations (such as calculating the intersection areas within each unit, line segments, and between units) are used to map point-like regions of interest, linear road networks, trajectories, and area remote sensing data to each target network unit, thus constructing a (multi)modal data set corresponding to each unit. This ensures that the modal data within each unit maintains consistency in spatial dimensions.

[0025] Furthermore, in order to alleviate sample scarcity and reduce annotation costs, this invention further divides the modality data set within each unit into a labeled dataset (containing the true value of the target indicator) and an unlabeled dataset (excluding the true value of the target indicator), so as to provide a data foundation for subsequent semi-supervised collaborative training.

[0026] Step S20: Input the modality dataset into a deep feature encoder for feature extraction, output physical features, functional features, and structural features, and map the physical features, functional features, and structural features to a unified feature space to calculate the correlation weights of the physical features, functional features, and structural features. Then, use all the correlation weights to fuse the physical features, functional features, and structural features to obtain a city spatial representation vector. Use the city spatial representation vector to construct labeled datasets and unlabeled datasets.

[0027] To comprehensively capture the multidimensional attributes of urban units, this invention discloses a deep feature encoder, such as... Figure 2 As shown, it includes: a visual environment encoding branch (Vision Transformer, ViT), a text semantic encoding branch (Bidirectional Encoder Representation of Transformer, BERT), and a spatial structure encoding branch (Graph Neural Network, GNN); each branch extracts features from data of different modalities to characterize the city's physical visual environment, social functional semantics, and spatial topology.

[0028] Specifically, the image dataset is input into the visual environment coding branch, various image data are filtered out, all the image data are divided into multiple image blocks, all the image blocks are mapped into high-dimensional embedding vectors, the position encoding of each cell is introduced into the high-dimensional embedding vectors, and the global features of the high-dimensional embedding vectors are extracted using a multi-layer attention mechanism to output physical features. The text dataset is input into the text semantic encoding branch, various types of text data are filtered out, all word units in the text data are mapped into comprehensive word vectors, and the comprehensive word vectors are analyzed using a bidirectional encoding structure to output functional features; The structure dataset is input into the spatial structure encoding branch, various spatial data are filtered out, a graph structure is constructed based on all the spatial data, and the graph structure is analyzed using graph convolutional network and graph attention network respectively, and structural features are output.

[0029] The visual environment coding branch is used to model street view images or remote sensing images corresponding to urban units. First, the input modal data is divided into image patches of fixed size. Then, a high-dimensional embedding vector is generated through linear mapping, while the identifier code of the target grid unit is introduced to preserve spatial structure information. Further, a multi-layer attention mechanism is used to extract the contextual features of the high-dimensional embedding vector. Finally, the physical features (i.e., visual feature representation) of the target region can be obtained through pooling operations.

[0030] Furthermore, the text semantic encoding branch models the semantic information of region of interest data, social media text, and other urban functions; through word segmentation and embedding layers, it maps lexical units into a comprehensive representation that integrates word embedding, position embedding, and paragraph embedding, and uses a bidirectional encoder structure to capture deep semantic relationships, thereby mining the combination patterns of regions of interest and semantic features of human activities within the region, and finally outputting functional features (i.e., text feature vectors).

[0031] Furthermore, the spatial structure encoding branch models the urban road network structure, trajectory flow relationships, or spatial adjacency relationships. First, it constructs a graph structure with spatial units as nodes and spatial adjacency or traffic connectivity relationships as edges, and assigns initial attribute features to the nodes. Then, it uses a graph convolutional network to perform multi-layer message passing and aggregation to capture the spatial topology and neighborhood dependencies, and finally outputs structural features (i.e., structural feature representation).

[0032] The physical features, functional features, and structural features are mapped to a unified feature space to obtain the corresponding transformation features; A certain transformation feature is randomly defined as the query vector, and the other two transformation features are defined as the key vector and value vector, respectively, to calculate the relevance weights between different transformation features: ; in, This indicates that functional characteristics are modeled using physical features. Indicates physical characteristics, The function represents the feature, and softmax represents the probability transformation function. This represents a query matrix transformed by physical and functional characteristics. The bond matrix represents the transformation between physical and functional characteristics. Indicates transpose. This represents a value matrix transformed by physical and functional characteristics. Indicates the feature dimension; In the adaptive gated fusion layer, all modeled feature representations are concatenated along the channel dimension, and each feature representation is fused according to all the aforementioned relevance weights to output a city spatial representation vector: ; in, Represents the urban spatial representation vector. This indicates an adaptive gating fusion layer. This represents the adaptive gated fusion function. This represents the physical feature representation after modeling. This represents the functional features after modeling. This represents the structural features after modeling. Indicates to , and Perform fusion processing; An initial labeled dataset is constructed using the urban spatial representation vector and the label, and the unlabeled data is constructed into an unlabeled dataset.

[0033] The multimodal data, after encoding and fusion, yields urban spatial representation vectors (covering all units), not all of which are unlabeled datasets. Labeled datasets contain units with urban spatial representation vectors and corresponding target city indicator labels, while unlabeled datasets contain units with urban spatial representation vectors but no corresponding labels.

[0034] Specifically, for the feature vectors output by the three-branch encoder, each modal feature is further mapped to a feature space of a unified dimension as the input representation for subsequent attention calculation. At the same time, a cross-attention mechanism is introduced, using a certain transformation feature as the query vector and defining the other two transformation features as the key vector and value vector, dynamically calculating the correlation weights between different modalities, thereby achieving deep alignment and information complementarity between different modal features, jointly modeling visual, text and structural features, and finally generating a unified feature representation.

[0035] The weights are calculated by measuring the scaled dot product similarity between feature vectors of different modalities (i.e., ...). If physical features and functional features exhibit a higher intrinsic correlation (high similarity) within the same grid cell, the system automatically determines that the combination contains more effective information and assigns it a higher weight.

[0036] Furthermore, an adaptive gating fusion layer is introduced into each feature representation to splice them along the channel dimension. The contribution ratio of different modal features is adjusted by gating weights. After nonlinear mapping and layer normalization, a unified urban spatial representation vector is output.

[0037] Step S30: Use the labeled dataset and the unlabeled dataset to perform alternating iterative training on the constructed deep neural network model and deep ensemble tree model, input the unlabeled dataset into the current deep neural network model and deep ensemble tree model, and output the first pseudo-label set and the second pseudo-label set.

[0038] In the context of limited labeled samples, this invention discloses a semi-supervised learning collaborative computing framework that achieves high-precision prediction of multi-dimensional urban indicators through a heterogeneous model collaborative training mechanism, dynamic pseudo-label generation, and multi-constraint joint screening mechanism.

[0039] Specifically, construct the initial deep neural network model and the initial deep ensemble tree model: ; ; in, This represents the initial deep neural network model. This represents the initial deep ensemble tree model. This represents the prediction error loss function. Represents a deep neural network model. Represents a deep ensemble tree model. Represents the first in the labeled dataset Urban spatial representation vector of each sample Indicates the first The label value of the target city indicator for each sample. This represents a labeled dataset. This indicates determining the optimal parameter settings for the deep neural network model. This indicates the optimal parameter settings for the deep ensemble tree model. In the current training epoch, the unlabeled dataset is input into the initial deep neural network model and the initial deep ensemble tree model for prediction, respectively, and outputs a first pseudo-label set and a second pseudo-label set: ; in, and These represent the first set of pseudo-tags and the second set of pseudo-tags, respectively. and They represent the first Deep neural network models and deep ensemble tree models trained in multiple iterations. Indicates the first in the unlabeled dataset One sample.

[0040] In the embodiments disclosed in this invention, two computational models with significantly different generalization abilities and learning preferences are integrated: a model based on a deep neural network (Multilayer Perceptron (MLP) or encoder, including but not limited to) and a model based on a deep ensemble tree model (which may utilize XGBoost (eXtremeGradient Boosting) in this invention, including but not limited to). Deep neural networks excel at capturing subtle nonlinear mapping patterns in high-dimensional feature spaces, while the deep ensemble tree model leverages its advantage in decision boundary segmentation on structured data to provide more stable baseline predictions.

[0041] In the training process disclosed in this invention, the deep neural network and deep ensemble tree model are initialized and supervised training is performed using an already labeled sample set (including labeled datasets and urban spatial representation vectors) to ensure that the two models acquire basic predictive capabilities. During the semi-supervised training in round t, the current deep neural network and deep ensemble tree model are used to predict the unlabeled dataset, generating a first pseudo-label set and a second pseudo-label set corresponding to the two models. In order to prevent erroneous pseudo-labels from introducing a negative transfer effect on model training, this invention uses a multi-constraint joint screening mechanism to include only high-confidence pseudo-label samples in the training set, and comprehensively evaluates the quality of pseudo-labels from multiple constraints (i.e., the multiple confidence verification constraints described below) to generate pseudo-labels using a deep neural network.

[0042] This invention introduces a semi-supervised learning mechanism, enabling the model to fully mine the potential structural information in massive unlabeled urban data with only a very small number of labeled samples, thus achieving stable and high-precision calculation and prediction of multidimensional urban indicators.

[0043] Step S40: Construct multiple reset confidence verification constraints, use all the confidence verification constraints to filter the target pseudo-label sample set from the first pseudo-label set and the second pseudo-label set, and add multiple pseudo-labels from the target pseudo-label sample set to the labeled dataset to obtain the target labeled dataset.

[0044] The confidence verification constraints include: stability constraints, model consistency constraints, distribution rationality constraints, and neighborhood consistency constraints.

[0045] Specifically, for the current training epoch, the stability constraints of the deep neural network model under multiple random perturbations are calculated: ; in, Indicates the first pseudo-label in the set. Stability constraints on the prediction variance of a sample under multiple random perturbations. Indicates the first The first set of pseudo-labels under random perturbation Indicates random perturbation; The model consistency constraints between the current deep neural network model and the current deep ensemble tree model are calculated using the first pseudo-label set and the second pseudo-label set: ; in, This represents the absolute value of the difference between the pseudo-labels output by the current deep neural network model and the current deep ensemble tree model. Indicates the consistency threshold; Obtain the statistical distribution characteristics of the target metrics in the labeled dataset, and use these characteristics to constrain the value range of the pseudo-labels: ; in, Indicates the first Pseudo-labels for each sample This represents the lower quantile of the labeled sample distribution. Indicates the high quantile of the labeled sample distribution; All samples are identified in the dual neighborhoods of the target unit's geographic space and feature space, resulting in a dual neighborhood set. The pseudo-labels of this dual neighborhood set maintain consistency with the mean of the known labels within the neighborhood. ; in, This represents a neighborhood consistency constraint. Indicates the first The neighborhood set of each sample Represents the first in the neighborhood set One sample, Represents the first in the neighborhood set The label corresponding to each sample Indicates the neighborhood consistency threshold; Based on the stability constraint, the model consistency constraint, the distribution rationality constraint, and the neighborhood consistency constraint, multiple pseudo-label samples are selected from the pseudo-label set to obtain the target pseudo-label sample set; The target pseudo-label sample set, the first pseudo-label set, and the pseudo-labels of the second pseudo-label are added to the annotation dataset to obtain the target annotation dataset: ; in, Indicates the first The labeled dataset for each training round. Indicates the first The labeled dataset for each training round. This represents the target pseudo-label sample set.

[0046] In this study, the prediction variance of the computational model (referring to a deep neural network in the current training epoch and a deep ensemble tree model in the next training epoch) under multiple random perturbations serves as a prediction stability constraint, retaining samples with highly stable results and excluding regions where the model's judgment is ambiguous. A model consistency constraint is constructed to ensure that the absolute difference between the predictions of two models for the same unit is within a low quantile range, utilizing the consensus between heterogeneous models to offset the inductive bias of a single model. A distribution rationality constraint is constructed, constraining the range of pseudo-label values ​​based on the statistical distribution characteristics of the target indicators in the labeled sample set, requiring them to fall within a reasonable range (defined as 20%-80% quantile in this invention) to avoid extreme outliers interfering with training. Finally, a neighborhood consistency constraint is defined, using the KNN algorithm (K-Nearest Neighbors) to find a dual neighborhood set of samples in both geographical coordinates and feature space, requiring that the pseudo-label maintain similarity to the mean of known labels or high-confidence pseudo-labels within the neighborhood.

[0047] A sample set that can simultaneously satisfy these four constraints is defined as a high-confidence pseudo-label sample set (i.e., the target pseudo-label sample set in this invention). The target pseudo-label sample set and its corresponding pseudo-label pair annotation dataset are expanded using a deep ensemble tree model (or a deep neural network model in the next round) to obtain an updated training dataset. This dataset is then used to update the parameters of the deep ensemble tree model (or a deep neural network model in the next round), thereby enabling the guided training of the deep ensemble tree model through the deep neural network model.

[0048] In this invention, a multi-constraint joint pseudo-label screening mechanism is used to strictly screen pseudo-labels of unlabeled samples from four dimensions: prediction stability, model consistency, distribution rationality, and spatial neighborhood consistency. This effectively suppresses the negative transfer risk caused by the introduction of erroneous pseudo-labels, thereby achieving stable convergence and performance improvement in the semi-supervised learning process.

[0049] Step S50: Train the current deep ensemble tree model using the target annotation dataset, and train the deep neural network model in the next round of training until a target prediction model is obtained. Then, use the target prediction model to predict the target region to be tested and output prediction results of multiple indicators.

[0050] Specifically, the current deep ensemble tree model is trained using the target-labeled dataset to update its parameters, thereby obtaining the deep ensemble tree model for the next training round. ; in, Indicates the first A deep ensemble tree model with training epochs, Represents a deep ensemble tree model. Indicates the first The training dataset for each training round. This represents the urban spatial representation vector of a sample in the target labeled dataset. This represents the label values ​​of samples in the target labeled dataset. This represents the prediction error loss function; The deep ensemble tree model from the next training round is used to assist the deep neural network model in the current round in iterative training until the target prediction model is obtained. The target prediction model is used to predict the unlabeled dataset and output the prediction results of multiple indicators for the target region.

[0051] If the training results of the current round do not yield a target prediction model that meets the requirements, a deep neural network model can be used to train the deep ensemble tree model. This alternating process is used for iterative training until the model performance converges or reaches the preset iteration limit. This alternating generation and optimization process enables the model to learn the intrinsic characteristics of the city from a large amount of unlabeled space, and finally output accurate urban indicator calculation and indicator prediction results.

[0052] Among them, a heterogeneous model collaborative computing framework is used to train deep neural network models and ensemble tree models together. Through complementary learning between the two types of models, the problem of bias accumulation and instability that a single model is prone to in small sample scenarios is effectively alleviated, and the generalization ability and robustness of semi-supervised learning models are significantly improved.

[0053] This invention can significantly reduce the dependence on labeled data and achieve high-precision labeled prediction across multiple city indicator calculation dimensions.

[0054] Furthermore, such as Figure 3 As shown, based on the above-mentioned method for predicting small-sample multi-dimensional urban indicators using semi-supervised learning, this invention also provides a system for predicting small-sample multi-dimensional urban indicators using semi-supervised learning, wherein the system includes: The data acquisition module 51 is used to acquire multiple modal data of the target area and multiple target city indicator data, and map all the modal data to spatial units to obtain a modal dataset, wherein the target city indicator data is small sample data and is used as labels to participate in training; The multimodal fusion module 52 is used to input the modal dataset into a deep feature encoder for feature extraction, output physical features, functional features and structural features, map the physical features, functional features and structural features to a unified feature space, calculate the correlation weights of the physical features, functional features and structural features, and use all the correlation weights to fuse the physical features, functional features and structural features to obtain a city spatial representation vector, and use the city spatial representation vector to construct a labeled dataset and an unlabeled dataset; The pseudo-label generation module 3 is used to perform alternating iterative training on the constructed deep neural network model and deep ensemble tree model using the labeled dataset and the unlabeled dataset, inputting the unlabeled dataset into the current deep neural network model and the deep ensemble tree model, and outputting a first pseudo-label set and a second pseudo-label set; The confidence verification module 54 is used to construct multiple reset confidence verification constraints, use all the confidence verification constraints to filter the target pseudo-label sample set from the first pseudo-label set and the second pseudo-label set, and add multiple pseudo-labels from the target pseudo-label sample set to the labeled dataset to obtain the target labeled dataset; The task output module 55 is used to train the current deep ensemble tree model using the target annotation dataset, train the deep neural network model in the next round of training until a target prediction model is obtained, and use the target prediction model to predict the target region to be tested, and output prediction results of multiple indicators.

[0055] Furthermore, such as Figure 4 As shown, based on the above-mentioned method and system for predicting small-sample multi-dimensional urban indicators based on semi-supervised learning, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 4 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0056] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a small-sample multi-dimensional city indicator prediction program 40 based on semi-supervised learning. This small-sample multi-dimensional city indicator prediction program 40 based on semi-supervised learning can be executed by the processor 10, thereby implementing the small-sample multi-dimensional city indicator prediction method based on semi-supervised learning in this application.

[0057] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the small-sample multi-dimensional urban indicator prediction method based on semi-supervised learning.

[0058] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components of the terminal communicate with each other via a system bus.

[0059] In one embodiment, when the processor 10 executes the semi-supervised learning-based small sample multidimensional city indicator prediction program 40 in the memory 20, it implements the steps of the semi-supervised learning-based small sample multidimensional city indicator prediction method as described above.

[0060] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a small sample multi-dimensional urban indicator prediction program based on semi-supervised learning, and when the small sample multi-dimensional urban indicator prediction program based on semi-supervised learning is executed by a processor, it implements the steps of the small sample multi-dimensional urban indicator prediction method based on semi-supervised learning as described above.

[0061] In summary, this invention provides a method and related equipment for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning. The method includes: acquiring multiple modal data of a target region and multiple target urban indicator data; mapping all the modal data to spatial units to obtain a modal dataset, wherein the target urban indicator data is small-sample data and serves as labels for training; inputting the modal dataset into a deep feature encoder for feature extraction, outputting physical features, functional features, and structural features, and fusing the physical features, functional features, and structural features to obtain an unlabeled dataset; and using the labeled dataset and the unlabeled dataset to train a constructed deep neural network and a deep... The deep ensemble tree model undergoes alternating iterative training. The unlabeled dataset is input into the current deep neural network and the deep ensemble tree model, outputting a corresponding set of pseudo-labels. Multiple confidence verification constraints are constructed, and multiple pseudo-label samples are selected from the pseudo-label set using all these constraints. The labeled dataset is updated using all the pseudo-label samples to obtain the target labeled dataset. The current deep ensemble tree model is trained using the target labeled dataset, and the deep neural network is trained in the next training round until a target prediction model is obtained. This model is then used to predict the target region under test, outputting prediction results for multiple indicators. This invention significantly reduces the dependence on labeled data and achieves high-precision labeled prediction across multiple city indicator calculation dimensions.

[0062] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0063] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0064] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning, characterized in that, The method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning includes: Multiple modal data of the target area and multiple target city indicator data are obtained, and all the modal data are mapped to spatial units to obtain a modal dataset, wherein the target city indicator data is small sample data and is used as labels for training; The modal dataset is input into a deep feature encoder for feature extraction, outputting physical features, functional features, and structural features. The physical features, functional features, and structural features are mapped to a unified feature space to calculate the correlation weights of the physical features, functional features, and structural features. The physical features, functional features, and structural features are then fused using all the correlation weights to obtain a city spatial representation vector. The city spatial representation vector is used to construct labeled and unlabeled datasets. The constructed deep neural network model and deep ensemble tree model are trained iteratively using the labeled dataset and the unlabeled dataset. The unlabeled dataset is input into the current deep neural network model and the deep ensemble tree model, and the first pseudo-label set and the second pseudo-label set are output. Construct multiple reset confidence verification constraints, use all the confidence verification constraints to filter the target pseudo-label sample set from the first pseudo-label set and the second pseudo-label set, and add multiple pseudo-labels from the target pseudo-label sample set to the labeled dataset to obtain the target labeled dataset; The current deep ensemble tree model is trained using the target labeled dataset, and the deep neural network model is trained in the next round of training until a target prediction model is obtained. The target prediction model is then used to predict the target region to be tested, and multiple indicator prediction results are output.

2. The method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning according to claim 1, characterized in that, The process of acquiring multiple modal data of the target area and multiple target city indicator data, and mapping all the modal data to spatial units to obtain a modal dataset, specifically includes: Acquire multiple modal data of the target region, and project each modal data into the constructed target specification network; The target specification network is queried by a spatial query algorithm, and all modal data in each unit are defined as a modal dataset.

3. The method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning according to claim 1, characterized in that, The deep feature encoder includes: a visual environment coding branch, a text semantic coding branch, and a spatial structure coding branch; The modal dataset includes: image dataset, text dataset, and structure dataset; The process of inputting the modality dataset into a deep feature encoder for feature extraction, and outputting physical features, functional features, and structural features, specifically includes: The image dataset is input into the visual environment coding branch, various image data are filtered out, all image data are divided into multiple image blocks, all image blocks are mapped into high-dimensional embedding vectors, the position encoding of each cell is introduced into the high-dimensional embedding vectors, and the global features of the high-dimensional embedding vectors are extracted using a multi-layer attention mechanism to output physical features. The text dataset is input into the text semantic encoding branch, various types of text data are filtered out, all word units in the text data are mapped into comprehensive word vectors, and the comprehensive word vectors are analyzed using a bidirectional encoding structure to output functional features; The structure dataset is input into the spatial structure encoding branch, various spatial data are filtered out, a graph structure is constructed based on all the spatial data, and the graph structure is analyzed using graph convolutional network and graph attention network respectively, and structural features are output.

4. The method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning according to claim 1, characterized in that, The process involves mapping the physical features, functional features, and structural features to a unified feature space, calculating the correlation weights of the physical features, functional features, and structural features, and then fusing these correlation weights to obtain a city spatial representation vector. This city spatial representation vector is then used to construct labeled and unlabeled datasets. Specifically, this includes: The physical features, functional features, and structural features are mapped to a unified feature space to obtain the corresponding transformation features; A certain transformation feature is randomly defined as the query vector, and the other two transformation features are defined as the key vector and value vector, respectively, to calculate the relevance weights between different transformation features: ； in, This indicates that functional characteristics are modeled using physical features. Indicates physical characteristics, The function represents the feature, and softmax represents the probability transformation function. This represents a query matrix transformed by physical and functional characteristics. The bond matrix represents the transformation between physical and functional characteristics. Indicates transpose. This represents a value matrix transformed by physical and functional characteristics. Indicates the feature dimension; In the adaptive gated fusion layer, all modeled feature representations are concatenated along the channel dimension, and each feature representation is fused according to all the aforementioned relevance weights to output a city spatial representation vector: ； in, Represents the urban spatial representation vector. This indicates an adaptive gating fusion layer. This represents the adaptive gated fusion function. This represents the physical feature representation after modeling. This represents the functional features after modeling. This represents the structural features after modeling. Indicates to , and Perform fusion processing; An initial labeled dataset is constructed using the urban spatial representation vector and the label, and the unlabeled data is constructed into an unlabeled dataset.

5. The method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning according to claim 1, characterized in that, The step of iteratively training the constructed deep neural network model and deep ensemble tree model using the labeled dataset and the unlabeled dataset, inputting the unlabeled dataset into the current deep neural network model and deep ensemble tree model, and outputting a first pseudo-label set and a second pseudo-label set, specifically includes: Constructing the initial deep neural network model and the initial deep ensemble tree model: ； ； in, This represents the initial deep neural network model. This represents the initial deep ensemble tree model. This represents the prediction error loss function. Represents a deep neural network model. Represents a deep ensemble tree model. Represents the first in the labeled dataset Urban spatial representation vector of each sample Indicates the first The label value of the target city indicator for each sample. This represents a labeled dataset. This indicates determining the optimal parameter settings for the deep neural network model. This indicates the optimal parameter settings for the deep ensemble tree model. In the current training epoch, the unlabeled dataset is input into the initial deep neural network model and the initial deep ensemble tree model for prediction, respectively, and outputs a first pseudo-label set and a second pseudo-label set: ； in, and These represent the first set of pseudo-tags and the second set of pseudo-tags, respectively. and They represent the first Deep neural network models and deep ensemble tree models trained in multiple iterations. Indicates the first in the unlabeled dataset One sample.

6. The method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning according to claim 5, characterized in that, The confidence verification constraints include: stability constraints, model consistency constraints, distribution rationality constraints, and neighborhood consistency constraints. The step of constructing multiple confidence verification constraints, using all the confidence verification constraints to filter the target pseudo-label sample set from the first pseudo-label set and the second pseudo-label set, and adding multiple pseudo-labels from the target pseudo-label sample set to the labeled dataset to obtain the target labeled dataset, specifically includes: For the current training epoch, calculate the stability constraints of the deep neural network model under multiple random perturbations: ； in, Indicates the first pseudo-label in the set. Stability constraints on the prediction variance of a sample under multiple random perturbations. Indicates the first The first set of pseudo-labels under random perturbation Indicates random perturbation; The model consistency constraints between the current deep neural network model and the current deep ensemble tree model are calculated using the first pseudo-label set and the second pseudo-label set: ； in, This represents the absolute value of the difference between the pseudo-labels output by the current deep neural network model and the current deep ensemble tree model. Indicates the consistency threshold; Obtain the statistical distribution characteristics of the target metrics in the labeled dataset, and use these characteristics to constrain the value range of the pseudo-labels: ； in, Indicates the first Pseudo-labels for each sample This represents the lower quantile of the labeled sample distribution. Indicates the high quantile of the labeled sample distribution; All samples are identified in the dual neighborhoods of the target unit's geographic space and feature space, resulting in a dual neighborhood set. The pseudo-labels of this dual neighborhood set maintain consistency with the mean of the known labels within the neighborhood. ； in, This represents a neighborhood consistency constraint. Indicates the first The neighborhood set of each sample Represents the first in the neighborhood set One sample, Represents the first in the neighborhood set The label corresponding to each sample Indicates the neighborhood consistency threshold; Based on the stability constraint, the model consistency constraint, the distribution rationality constraint, and the neighborhood consistency constraint, multiple pseudo-label samples are selected from the pseudo-label set to obtain the target pseudo-label sample set; The target pseudo-label sample set, the first pseudo-label set, and the pseudo-labels of the second pseudo-label are added to the annotation dataset to obtain the target annotation dataset: ； in, Indicates the first The labeled dataset for each training round. Indicates the first The labeled dataset for each training round. This represents the target pseudo-label sample set.

7. The method for predicting small-sample, multi-dimensional urban indicators based on semi-supervised learning according to claim 1, characterized in that, The process involves training the current deep ensemble tree model using the target-labeled dataset, training the next deep neural network model in the next training round until a target prediction model is obtained, and then using the target prediction model to predict the target region to be tested, outputting prediction results for multiple indicators, specifically including: The current deep ensemble tree model is trained using the target-labeled dataset to update its parameters, resulting in the deep ensemble tree model for the next training epoch. ； in, Indicates the first A deep ensemble tree model with training epochs, Represents a deep ensemble tree model. Indicates the first The training dataset for each training round. This represents the urban spatial representation vector of a sample in the target labeled dataset. This represents the label values ​​of samples in the target labeled dataset. This represents the prediction error loss function; The deep ensemble tree model from the next training round is used to assist the deep neural network model in the current round in iterative training until the target prediction model is obtained. The target prediction model is used to predict the unlabeled dataset and output the prediction results of multiple indicators for the target region.

8. A small-sample, multi-dimensional urban indicator prediction system based on semi-supervised learning, characterized in that, The semi-supervised learning-based small-sample multi-dimensional city indicator prediction system is used to implement the semi-supervised learning-based small-sample multi-dimensional city indicator prediction method as described in any one of claims 1-7, wherein the semi-supervised learning-based small-sample multi-dimensional city indicator prediction system comprises: The data acquisition module is used to acquire multiple modal data of the target area and multiple target city indicator data, and map all the modal data to spatial units to obtain a modal dataset, wherein the target city indicator data is small sample data and is used as labels to participate in training; The multimodal fusion module is used to input the modal dataset into a deep feature encoder for feature extraction, output physical features, functional features, and structural features, and map the physical features, functional features, and structural features to a unified feature space to calculate the correlation weights of the physical features, functional features, and structural features. The module then uses all the correlation weights to fuse the physical features, functional features, and structural features to obtain a city spatial representation vector. The module uses the city spatial representation vector to construct labeled and unlabeled datasets. The pseudo-label generation module is used to perform alternating iterative training on the constructed deep neural network model and deep ensemble tree model using the labeled dataset and the unlabeled dataset, inputting the unlabeled dataset into the current deep neural network model and the deep ensemble tree model, and outputting a first pseudo-label set and a second pseudo-label set; The confidence verification module is used to construct multiple reset confidence verification constraints, use all the confidence verification constraints to filter the target pseudo-label sample set from the first pseudo-label set and the second pseudo-label set, and add multiple pseudo-labels from the target pseudo-label sample set to the labeled dataset to obtain the target labeled dataset; The task output module is used to train the current deep ensemble tree model using the target annotation dataset, train the deep neural network model in the next round of training, until a target prediction model is obtained, and use the target prediction model to predict the target region to be tested, and output prediction results of multiple indicators.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a small-sample multi-dimensional city indicator prediction program based on semi-supervised learning, which is stored in the memory and can run on the processor. When the small-sample multi-dimensional city indicator prediction program based on semi-supervised learning is executed by the processor, it implements the steps of the small-sample multi-dimensional city indicator prediction method based on semi-supervised learning as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a small-sample multi-dimensional city indicator prediction program based on semi-supervised learning, which, when executed by a processor, implements the steps of the small-sample multi-dimensional city indicator prediction method based on semi-supervised learning as described in any one of claims 1-7.

Images